jdelony/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-06-28 04:30:15 -05:00
Code Issues Packages Projects Releases Wiki Activity

Fix Mistral3 split mode graph

Browse Source
This commit is contained in:
Kawrakow 2026-05-10 05:46:40 +00:00
parent 4bbdb8ed0b
commit e7f8d7cdbd
1 changed files with 7 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
src/graphs/build_mistral3.cpp
View File

@ -25,42 +25,31 @@ ggml_cgraph * llm_build_context::build_mistral3() {
ggml_tensor * KQ_mask = build_inp_KQ_mask();
ggml_tensor * inp_out_ids = build_inp_out_ids();
ggml_tensor * inp_out_ids = n_tokens > 1 ? build_inp_out_ids() : nullptr;
//const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : 1.f;
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL;
auto rope_factors = build_rope_factors(il);
cur = build_std_attention(gf, model.layers[il].attn_norm, inpL, inp_pos, nullptr, rope_factors, KQ_mask,
nullptr, inp_attn_scale, kq_scale, hparams.f_attention_scale, 0, il);
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
cb(cur, "last_attn", il);
cb(inpSA, "last_ffn_inp", il);
}
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
cur = build_std_attention(gf, model.layers[il].attn_norm, inpL, inp_pos, il == n_layer - 1 ? inp_out_ids : nullptr,
rope_factors, KQ_mask, nullptr, inp_attn_scale, kq_scale, hparams.f_attention_scale, 0, il, true, false, true);
// feed-forward network (non-MoE)
if (model.layers[il].ffn_gate_inp == nullptr) {
// non-MoE
cur = llm_build_ffn(ctx0, lctx, model.layers[il].ffn_norm, ffn_inp,
cur = llm_build_ffn(ctx0, lctx, model.layers[il].ffn_norm, cur,
model.layers[il].ffn_up, model.layers[il].ffn_up_b, nullptr,
model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, nullptr,
model.layers[il].ffn_down, model.layers[il].ffn_down_b, nullptr,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, cb, il, gf);
LLM_FFN_SILU, LLM_FFN_PAR, cb, il, gf, true);
cb(cur, "ffn_out", il);
} else {
// MoE branch
cur = llm_build_std_moe_ffn(ctx0, lctx, model.layers[il].ffn_norm, ffn_inp,
cur = llm_build_std_moe_ffn(ctx0, lctx, model.layers[il].ffn_norm, cur,
model.layers[il].ffn_gate_inp, nullptr,
model.layers[il].ffn_up_exps, nullptr,
model.layers[il].ffn_gate_exps, nullptr,
@ -72,9 +61,8 @@ ggml_cgraph * llm_build_context::build_mistral3() {
n_expert, n_expert_used,
LLM_FFN_SILU, true, false, 0.0f,
LLM_EXPERT_GATING_FUNC_SOFTMAX,
LLM_FFN_SILU, cb, il, gf);
LLM_FFN_SILU, cb, il, gf, true);
}
cur = ggml_add(ctx0, cur, ffn_inp);
cb(cur, "ffn_out", il);
cur = lctx.cvec.apply_to(ctx0, cur, il);